What is document intelligence?
Document intelligence is the automated extraction of structure, meaning, and relationships from unstructured and semi-structured documents — PDFs, scanned images, drawings, spreadsheets, contracts, invoices, emails — so that the information inside them becomes queryable, auditable, and ready to feed downstream agents.
EXXY® AI Studio treats document intelligence as foundational infrastructure. Every other capability — presentations, BIM analysis, specifications, meeting minutes — depends on a clean, citable, indexed view of source material.
What the document intelligence layer does
1. Ingestion across formats
- Native digital PDFs.
- Scanned PDFs and image-only documents (OCR with layout preservation).
- DWG/DXF drawings and PDF plot sheets.
- Office formats (DOCX, XLSX, PPTX, ODT).
- Email archives (MSG, EML).
- Plain text, Markdown, CSV, JSON.
- BIM model exports (IFC; cross-referenced with the BIM analysis agent).
2. Layout-aware parsing
Documents are parsed with a layout model so columns, headers, footers, tables, drawings, and call-outs are preserved. This avoids the classic OCR-without-layout problem where a multi-column page becomes a jumble.
3. Entity and field extraction
Configurable extractors pull named fields ("project number", "section reference", "specified product", "monetary amount", "due date") into structured records with confidence scores.
4. Cross-document relationships
The collation agent links related items across the corpus: an action item from a meeting transcript → the spec section it references → the BIM element it affects → the prior project where the same product was used.
5. Citable answers
Every answer the AI gives over your corpus comes with a citation that points to the exact page, paragraph, or cell. Click the citation and you land on the source.
Why citation-first matters
In a regulated profession, an AI answer without a citation is worthless. EXXY® AI Studio enforces citation by construction: the retrieval layer hands the model chunks with IDs, and the answer schema requires a citation array referencing those IDs. If the model produces an unsupported claim, the reviewer agent rejects the answer.
This pattern eliminates the most common failure mode of naive RAG implementations — the model "remembering" something that was not actually in the retrieved context.
Use cases
- Tender response: Ingest the RFP, all addenda, and your office's prior responses; the agent drafts each section with citations to the RFP requirements.
- Project handover: Collate every drawing, specification, and meeting minute; the agent answers "what was decided about the lobby ceiling?" with the source.
- Due diligence: Pull dozens of contracts; the agent extracts term sheets and flags non-standard clauses.
- Knowledge continuity: When a senior staff member leaves, the documents they touched become a permanent, queryable knowledge base instead of a black box.
How collation works in practice
- Upload — drag a folder of mixed documents into the project.
- Parse — layout-aware extraction runs in the background; you see progress per file.
- Index — content is embedded into the project's vector store.
- Query — ask in natural language; answers come back with citations.
- Promote — promote any answer into a generated slide or drafted specification.
Privacy
- Documents are stored inside your tenant boundary.
- Vector embeddings are stored alongside the project, never shared across tenants.
- Source files and embeddings are excluded from any external model training. See our AI training preferences.
FAQ
How accurate is OCR on scanned drawings?
Layout-aware OCR achieves ~98% character accuracy on clean scans and ~92% on poor-quality scans. Tables and call-outs are reconstructed with explicit cell/region mapping rather than flattened text.
Can it read handwritten annotations?
Yes, with a separate handwriting recognition pass. Handwritten content is flagged with lower confidence and surfaced for human review.
Does it work with multi-language corpora?
Yes. English and Vietnamese are first-class. Mixed-language documents are handled per paragraph.
How are confidential documents protected?
All ingestion runs inside the tenant boundary. Documents marked confidential are encrypted at rest with project-scoped keys, and retrieval is authentication-checked on every query.
Can I purge a document and all its derived data?
Yes. Deletion is recursive: source file, parsed text, embeddings, and any cached AI answers grounded in that document are all purged. The deletion is logged in the audit trail.
Bring your corpus in
Open a project in EXXY® AI Studio, drop your documents, and start asking questions — or read about agentic orchestration to see how this layer powers every other capability.